wasmtime 49 bump: pin-drift xfail register after the CM#705 pin advance (FACT reentrance, name rules, max-value-size)

[lannbot]

After the CM pin advance to 2f13265 (CM#705 merge; polyengine#173), a bounded set of new conformance rows cannot pass on wasmtime-environ 47.0.3 and are xfailed under pin-drift classes. This issue is the register; the wasmtime bump (49.x, cf. the parked findings entries from #246/#247 and the case-folding gap #238) retires every class below. Failure direction is benign throughout: 47-era tooling over-rejects now-valid programs or under-rejects now-invalid ones at validation time — no soundness exposure at runtime.

Class: fact-reentrance-47 — FACT compiles now-valid reentrance to static traps

wasmtime-environ 47's FACT emits an unconditional CannotEnterComponent trap at translation time for same-instance/ancestor lower→lift pairs (upstream trampoline.rs; documented at runtime/src/intrinsics/fact_calls.ts header). CM#705 makes these calls valid. Affected: 10 of 13 directives in test/async/reentrance.wast (all route at least one executed adapter edge through an ancestor lift — including the sibling-reentrance cases 4/10/11, whose wiring passes through a root-level lift). The three passable directives (parent-to-child control, the never-executes-adapter branch, and the dtor-reentrance case) are NOT in this class — the dtor case is gated on polyengine's own gate removal (#173) instead.

Class: name-rules-47 — hyphen-insensitive strongly-unique names (CM#703/#704)

wasmparser at the 47 pin predates hyphen-insensitivity and transitive strongly-unique rules: 5 new assert_invalid directives in test/validation/kebab.wast are accepted (missing rejection). Extends the case-folding gap already tracked in #238 and the findings entries (#246).

Class: max-value-size-47 — CM#688 static size bound

test/validation/max-value-size.wast (new): 7 assert_invalid directives pinning elem_size(t) < 2^28, including product-overflow and u32-wrap cases. The 47-era validator accepts all 7. Note: this file is in upstream's own test/nyi.txt at the pin — even current wasmtime does not implement it; verify at bump time whether the bumped wasmparser enforces it or the class needs to survive the bump with a narrower name.

Exit criteria

• wasmtime-environ bump lands (translator-shim Cargo pin + just test-rust + full regen/triage).
• All three classes' xfail entries removed; the stale-xfail detector enforces completeness.
• translator: accepts [method]/[static] case-folded name collisions CM validation requires rejecting (inherited wasmparser gap) #238's translator-acceptance gap re-checked against the bumped wasmparser.

[lannbot]

Correction to the fact-reentrance-47 class description, from ground-truth triage during the pin advance (probe of async/reentrance.json + wasmtime-environ 47.0.3 source):

The static trap is broader than described. fact/trampoline.rs:116-127 traps unconditionally when lift.instance == lower.instance or either instance is an ancestor of the other — i.e. every vertical fused adapter is a compile-time trap stub at 47, including plain non-reentrant parent→child calls (a canon lower at the outer level of a child instance's lifted export). Only sibling adapters compile to real fused code, which is why the existing corpus (sibling/host-shaped composition) never surfaced this: under the old pin, the only corpus rows exercising vertical adapters were trap-on-reenter.wast-family assert_trap "cannot enter component instance" rows, and the static stub's message is byte-compatible with the dynamic check's — it passed for a subtly different reason than we thought.

Consequences:

• All 10 executed-adapter directives in reentrance.wast are in this class — including the two rows pre-analysis called "controls" (parent-to-child, wast line 60) and the dtor-reentrance row (line 198, predicted to be provable by polyengine's own gate removal). None of them reach polyengine runtime code: they die inside the generated adapter with wasm trap: cannot enter component instance (the wasm trap: prefix is the discriminator vs polyengine's own gate message, which carries the instance index and (reentrance forbidden)).
• Line 100 (f 0) is a poison cascade: line 99's adapter trap breaks the host-entry bracket, polyengine's per-instance poisoning marks the corpse, and line 100 is then refused by the poisoning gate with the Go components trap (and poison their instance) when the Go runtime reads the clock inside cabi_realloc #145 cause suffix naming the adapter trap. It clears when line 99 clears, so it rides this class too.
• The cm705-gate-removal xfail class (pin-advance migration map: CM#705 removes the may_enter flag/trap #173) has zero corpus rows: the gate-removal PR's proof burden shifts entirely to the runtime unit suites (inverted reentrance pins + poisoning pins), and post-removal these corpus rows keep failing identically until the wasmtime bump.
• Exit criterion sharpened: upstream wasmtime removed the static vertical-adapter trap while adopting CM#705, so the bump makes these adapters real fused code; expect the whole class (11 rows incl. the cascade) to flip at once. Lines 522/657 (+ their 645/760 cascades) are NOT this class — wasmparser 0.252 encoding drift, tracked with Pin-blocked conformance residues (#13 exit): name-option + thread-encoding xfails need wasmparser >= 0.255 #152.